#include "simgrid/s4u/Actor.hpp"
#include "src/simix/popping_private.h"
#include "xbt/swag.h"
-#include<list>
+#include <list>
typedef struct s_smx_process_exit_fun {
int_f_pvoid_pvoid_t fun;
void *arg;
-} s_smx_process_exit_fun_t, *smx_process_exit_fun_t;
+} s_smx_process_exit_fun_t;
+typedef s_smx_process_exit_fun_t* smx_process_exit_fun_t;
namespace simgrid {
namespace simix {
/* Refcounting */
private:
- std::atomic_int_fast32_t refcount_{1};
+ std::atomic_int_fast32_t refcount_{0};
+
public:
friend void intrusive_ptr_add_ref(ActorImpl* process)
{
- process->refcount_.fetch_add(1, std::memory_order_relaxed);
+ // std::memory_order_relaxed ought to be enough here instead of std::memory_order_seq_cst
+ // But then, we have a threading issue when an actor commits a suicide:
+ // it seems that in this case, the worker thread kills the last occurrence of the actor
+ // while usually, the maestro does so. FIXME: we should change how actors suicide
+ process->refcount_.fetch_add(1, std::memory_order_seq_cst);
}
friend void intrusive_ptr_release(ActorImpl* process)
{
// inspired from http://www.boost.org/doc/libs/1_55_0/doc/html/atomic/usage_examples.html
if (process->refcount_.fetch_sub(1, std::memory_order_release) == 1) {
+ // Make sure that any changes done on other threads before their acquire are committed before our delete
+ // http://stackoverflow.com/questions/27751025/why-is-an-acquire-barrier-needed-before-deleting-the-data-in-an-atomically-refer
std::atomic_thread_fence(std::memory_order_acquire);
delete process;
}
